Scholarlink

<p>The growing demand for urban parking has imposed increasingly stringent requirements on the real-time and fine-grained management of parking resources. Accurate parking availability (PA) forecasting is therefore essential for parking guidance and traffic congestion mitigation. Existing PA forecasting methods still face two fundamental challenges: first, non-stationary and multi-scale temporal structures are difficult to model; second, global spatial dependency modeling must balance predictive accuracy against computational efficiency. To address these challenges, this paper proposes DHPA, a decomposition-enhanced hybrid spatio-temporal forecasting framework built upon the DeepPA architecture. In the temporal dimension, DHPA introduces a temporal decomposition block, which explicitly models trend and periodic structures to reduce the difficulty of learning non-stationary temporal patterns, thereby improving the stability and generalization of multi-step forecasting. In the spatial dimension, DHPA introduces a hierarchical spatial learning block, which replaces full-node attention with approximate global interaction and enhances spatial dependency modeling under controlled computational cost. Main experiments on the SINPA dataset show that, over the overall 3-hour forecasting horizon, DHPA reduces the mean absolute error by 1.7% compared with the strongest baseline.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DHPA: decomposition-enhanced hybrid spatiotemporal modeling for city-scale parking availability forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Yufan Wang

摘要

The growing demand for urban parking has imposed increasingly stringent requirements on the real-time and fine-grained management of parking resources. Accurate parking availability (PA) forecasting is therefore essential for parking guidance and traffic congestion mitigation. Existing PA forecasting methods still face two fundamental challenges: first, non-stationary and multi-scale temporal structures are difficult to model; second, global spatial dependency modeling must balance predictive accuracy against computational efficiency. To address these challenges, this paper proposes DHPA, a decomposition-enhanced hybrid spatio-temporal forecasting framework built upon the DeepPA architecture. In the temporal dimension, DHPA introduces a temporal decomposition block, which explicitly models trend and periodic structures to reduce the difficulty of learning non-stationary temporal patterns, thereby improving the stability and generalization of multi-step forecasting. In the spatial dimension, DHPA introduces a hierarchical spatial learning block, which replaces full-node attention with approximate global interaction and enhances spatial dependency modeling under controlled computational cost. Main experiments on the SINPA dataset show that, over the overall 3-hour forecasting horizon, DHPA reduces the mean absolute error by 1.7% compared with the strongest baseline.